American Economic Association

Bad Beta, Good Beta Author(s): John Y. Campbell and Tuomo Vuolteenaho Source: The American Economic Review, Vol. 94, No. 5 (Dec., 2004), pp. 1249-1275 Published by: American Economic Association Stable URL: http://www.jstor.org/stable/3592822 Accessed: 30/03/2010 07:48
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/action/showPublisher?publisherCode=aea. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

American Economic Association is collaborating with JSTOR to digitize, preserve and extend access to The American Economic Review.

http://www.jstor.org

Bad Beta, Good Beta

AND TUOMOVUOLTEENAHO* By JOHN Y. CAMPBELL

This paper explains the size and value "anomalies" in stock returns using an economically motivated two-beta model. We break the beta of a stock with the marketportfolio into two components,one reflectingnews about the market'sfuture cash flows and one reflectingnews about the market'sdiscount rates. Intertemporal asset pricing theorysuggests that theformer should have a higherprice of risk; thus beta, like cholesterol, comes in "bad" and "good" varieties. Empirically,we find that value stocks and small stocks have considerably higher cash-flow betas than growth stocks and large stocks, and this can explain their higher average returns. The poor performance of the capital asset pricing model (CAPM) since 1963 is explained by the fact that growth stocks and high-past-betastocks have predominantly good betas with low risk prices. (JEL G12, G14, N22)

How should a rational investor measure the risks of stock marketinvestments?What determines the risk premium that will induce a rational investor to hold an individual stock at its market weight, rather than overweighting or underweightingit? According to the CAPM of William Sharpe (1964) and John Lintner (1965), a stock's risk is summarizedby its beta with the marketportfolio of all invested wealth. Controllingfor beta, no other characteristicsof a stock should influencethe returnrequiredby a rationalinvestor. It is well known that the CAPM fails to describe averagerealized stock returnssince the early 1960s, if a value-weightedequity index is used as a proxy for the market portfolio. In
* Campbell: Departmentof Economics, LittauerCenter, HarvardUniversity, Cambridge,MA 02138, and National Bureau of Economic Research (e-mail: john_campbell@ Vuolteenaho:Departmentof Economics, Litharvard.edu); tauer Center, HarvardUniversity, Cambridge,MA 02138, and NBER (e-mail: t_vuolteenaho@harvard.edu).We would like to thank Ben Bernanke, Michael Brennan, Joseph Chen, Randy Cohen, RobertHodrick,Matti Keloharju, Owen Lamont,Greg Mankiw,Lubos Pastor,Antti Petajisto, Christopher Polk, Jay Shanken, Andrei Shleifer, Jeremy Stein, Sam Thompson,Luis Viceira, two anonymousreferees, and seminar participantsat various venues for helpful comments. We are grateful to Ken French for providing us with some of the data used in this study. All errors and omissions remain our responsibility.This materialis based upon work supportedby the National Science Foundation under GrantNo. 0214061 to Campbell.

particular,small stocks and value stocks have deliveredhigher averagereturnsthantheirbetas can justify. Adding insult to injury, stocks with high past betas have had average returns no higher than stocks of the same size with low past betas. These findings tempt investors to tilt their stock portfolios systematically toward small stocks, value stocks, and stocks with low past betas.1 We arguethat returnson the marketportfolio have two components, and that recognizing the difference between these two components can eliminate the incentive to overweight value, small, and low-beta stocks. The value of the marketportfolio may fall because investors receive bad news about future cash flows; but it may also fall because investors increase the discountrate or cost of capitalthat they apply to these cash flows. In the first case, wealth decreases and investment opportunities are unchanged, while in the second case, wealth decreases but future investment opportunities
improve.

1 Seminal early referencesinclude Rolf Banz (1981) and Marc Reinganum (1981) for the size effect, and Benjamin Grahamand David Dodd (1934), Sanjoy Basu (1977, 1983), Ray Ball (1978), and Barr Rosenberg et al. (1985) for the value effect. Eugene Fama and KennethFrench(1992) give an influentialtreatmentof both effects within an integrated framework and show that sorting stocks on past market betas generates little variationin average returns.

1249

1250

THEAMERICAN ECONOMICREVIEW

DECEMBER2004

These two components should have different significance for a risk-averse, long-term investor who holds the market portfolio. Such an investor may demanda higher premiumto hold assets that covary with the market's cash-flow news than to hold assets that covary with news about the market's discount rates, for poor returns driven by increases in discount rates are partially compensated by improved prospects for future returns.To measure risk for this investor properly, the single beta of the SharpeLintner CAPM should be broken into two differentbetas: a cash-flow beta and a discountrate beta. We expect a rational investor who is holding the market portfolio to demand a greater reward for bearing the former type of risk than the latter. In fact, an intertemporal capital asset pricing model (ICAPM) of the sort proposed by Robert Merton (1973) suggests that the price of risk for the discount-ratebeta should equal the variance of the marketreturn, while the price of risk for the cash-flow beta should be y times greater,where y is the investor's coefficient of relative risk aversion. Thus, if the investor is conservative in the sense that y > 1, the cash-flow beta has a higher price of risk. An intuitive way to summarizeour story is to say that beta, like cholesterol, has a "bad"variety and a "good" variety. Just as a person's heart-attack is determinednot by his overall risk cholesterol level but primarilyby his bad cholesterol level with a secondary influence from good cholesterol, so the risk of a stock for a long-term investor is determined not by the stock's overall beta with the marketbut by its bad cash-flow beta with a secondary influence from its good discount-ratebeta. Of course, the good beta is good not in absolute terms but in relation to the other type of beta. We test these ideas by fitting a two-beta ICAPM to historical monthly returnson stock ratios, portfolios sortedby size, book-to-market and marketbetas. We consider not only a sample period since 1963 that has been the subject of much recent research, but also an earlier sample period 1929-1963 using the data of James Davis et al. (2000). In the moder period, July 1963 to December 2001, we find that the two-beta model greatly improves the poor performance of the standard CAPM. The main

reason for this is that growth stocks, with low averagereturns,have high betas with the market portfolio;but theirhigh betas are predominantly good betas, with low risk prices. Value stocks, with high averagereturns,have higherbad betas than growth stocks do. The two-beta model also explains why stocks with high past CAPM betas have offered relatively little extra return:these stocks have higher good betas but almost the same bad betas as other stocks. Since the good beta carriesonly a low premium,the almost flat relationbetween averagereturnsand the CAPM beta is no puzzle to the two-beta model. In the early period, January 1929 to June 1963, we find that the ratio of good to bad beta is relatively constantacross the assets we consider, so the single-beta CAPM adequately explains the data. Our model explains why stocks with high cash-flow betas may offer high average returns, given that long-terminvestors are fully invested in equities at all times, or, in a slight generalization of the model, maintaina constantallocation to equities. Our model does not explain why long-term investors would wish to keep their equity allocations constant. If the equity premium is time-varying, it is optimal for a long-term investor with a fixed coefficient of relative risk aversion to invest more in equities at times when the equity premium is high (Campbell and Luis Viceira, 1999; Tong Suk Kim and Edward Omberg, 1996). We could generalize the model to allow a time-varying equity weight in the investor's portfolio,but this would not be consistent with general equilibrium if all investors have the same preferences. Thus our model cannot be interpreted as a model representative-agent general-equilibrium of the economy. Our achievement is merely to show that the risks of value, small, and lowpast-beta stocks are sufficient to deter investment in these stocks by conservative long-term investors who eschew markettiming.2
2 There are numerous competing explanations for the size and value effects. The ArbitragePricing Theory (APT) of Stephen Ross (1976) allows any pervasive source of common variation to be a priced risk factor. Fama and French(1993) introducean influentialthree-factormodel to describe the size and value effects in average returns.Ravi and Zhenyu Wang (1996), Martin Lettau and Jagannathan

VOL.94 NO. 5

CAMPBELL AND VUOLTEENAHO: BAD BETA, GOOD BETA

1251

In developing and testing the two-beta ICAPM, we draw on a great deal of related literature.The idea that the market'sreturncan be attributed to cash-flow and discount-rate news is not novel. Campbelland Robert Shiller frame(1988a) develop a loglinear approximate work in which to study the effects of changing cash-flow and discount-rateforecasts on stock prices. Campbell (1991) uses this framework and a vector autoregressive (VAR) model to decompose marketreturnsinto cash-flow news and discount-rate news. Empirically, he finds thatdiscount-rate news is far from negligible; in U.S. data, for example, his VAR syspostwar tem explains most stock-returnvolatility as the result of discount-ratenews. The insight that long-term investors care about shocks to investmentopportunitiesis due to Merton (1973). Campbell (1993) solves a discrete-time empirical version of Merton's ICAPM, assuming that asset returns are homoskedastic and that a representativeinvestor has the recursivepreferencesproposedby Lawrence Epstein and Stanley Zin (1989, 1991). The solution is exact in the limit of continuous investor has elasticity time, if the representative of intertemporal substitutionequal to one, and is otherwise a loglinear approximation.Campbell writes the solution in the form of a K-factor model, where the firstfactoris the marketreturn

Sydney Ludvigson (2001), and Lu Zhang and Ralitsa Petkova (2002) arguethatthe CAPM might hold conditionally, but fail unconditionally, although JonathanLewellen and Stefan Nagel (2003) show that the magnitudeof the value effect is too large to be explainedby the conditionalCAPM. Tobias Adrianand FrancescoFranzoni(2004) and Lewellen and Jay Shanken (2002) explore learning as a possible explanation of these anomalies. Richard Roll (1977) emphasizes that tests of the CAPM are misspecified if one cannot measure the marketportfolio correctly. While Robert Stambaugh(1982) and Shanken(1987) find thatthe tests of the CAPM are insensitive to the inclusion of other financial assets, Campbell (1996), Jagannathanand Wang (1996), and Lettau and Ludvigson (2001) find that humancapital wealth may be important.Josef Lakonishok et al. (1994), Rafael La Porta (1996), and La Porta et al. (1997) argue that investors' irrationalitydrives the value effect. Alon Brav et al. (2002) show that analysts' price targets imply high subjective expected returns on growth stocks, consistent with the hypothesis that the value effect is due to expectationalerrors.

and the otherfactors are shocks to variablesthat predict the marketreturn.3 The two recent empiricalpapersthatare closest to ours in their focus are by Michael J. Brennanet al. (2004) and Joseph Chen (2003). Brennan et al. model the riskless interest rate and the Sharperatio on the marketportfolio as continuous-time AR(1) processes. They estimate the parametersof their model using bond market data and explore the model's implications for the value and size effects in U.S. equities since 1953, with some success. Chen (2003) extends the framework of Campbell (1993) to allow for heteroskedastic asset returns, but given the state variables he includes in his model, he finds little evidence that growth stocks are valuable hedges against shocks to investment opportunities. A key to our success in explaining a number of asset pricing anomalies is our use of the small-stock value spread to predict aggregate stock returns. Recently, several authors have found that high returnsto growth stocks, particularly small growth stocks, seem to forecastlow returnson the aggregate stock market. Venkat R. Eleswarapu and Reinganum (2004) use lagged three-yearreturnson an equal-weighted index of growth stocks, while Brennan et al. (2001) use the difference between the log bookto-market ratios of small growth stocks and small value stocks to predictthe aggregatemarket. In this paper we use a measure similar to that of Brennan et al. (2001) and find that, indeed, growth stock returnshave high covariances with declines in marketdiscount rates. It is naturalto ask why high returnson small growth stocks should predictlow returnson the stock market as a whole. This is a particularly importantquestion since time-series regressions of aggregatestock returnson arbitrary predictor variables can easily produce meaningless datamined results. The most powerful motivationis provided by the ICAPM itself. We know that value stocks outperformgrowth stocks, particularly among smaller stocks, and that this cannot be explainedby the traditionalstatic CAPM.
3 Campbell (1996), Yuming Li (1997), Robert Hodrick et al. (1999), Anthony Lynch (2001), Brennanet al. (2001, 2004), David Ng (2004), Hui Guo (2003), and Chen (2003) explorethe empiricalimplicationsof Merton'smodel.

1252

THEAMERICAN ECONOMICREVIEW

DECEMBER2004

If the ICAPM is to explain this anomaly, then small growth stocks must have intertemporal hedging value that offsets their low returns;that is, their returns must be negatively correlated In with innovationsto investmentopportunities. orderto evaluate this hypothesis it is naturalto ask whether a long moving average of smallgrowth-stockreturnspredictsinvestmentopportunities. This is exactly what we do when we include the small-stockvalue spreadin our forecasting model for marketreturns.In short, the small-stockvalue spreadis not an arbitrary forevariable,but one thatis suggestedby the casting asset pricing theory we are trying to test. The organizationof the paperis as follows. In Section I, we estimate two components of the return on the aggregate stock market, one caused by cash-flow shocks and the other by discount-rateshocks. In Section II, we use these componentsto estimatecash-flow and discountrate betas for portfolios sorted on firm characteristics and risk loadings. In Section III, we lay out the intertemporalasset pricing theory that justifies differentrisk premia for bad cash-flow beta and good discount-ratebeta. We also show that the returns to small and value stocks can largely be explained by allowing different risk premiafor these two differentbetas. Section IV concludes.
I. How Cash-Flow News and Discount-Rate News Move the Market

(1)

rt+I

E,r,t+ = (Et+1 Et) I

j=o

fAdt+1+j

- (Et+

- Et) E Pirt+ l+j

j=1

NCF,t+i

NDR,t + 1

A simple present-valueformulapoints to two reasons why stock prices may change. Either expected cash flows change, discount rates change, or both. In this section, we empirically estimate these two components of unexpected returnfor a value-weighted stock marketindex. Consistent with findings of Campbell (1991), the fitted values suggest that over our sample period (January 1929 to December 2001) discount-rate news causes much more variation in monthly stock returns than cash-flow news. A. Return-Decomposition Framework Following Campbell and Shiller (1988a) and Campbell (1991), we use a loglinear approximate decomposition of returns:

where r + is a log stock return,d+ 1 is the log dividend paid by the stock, A denotes a oneperiod change, Et denotes a rationalexpectation at time t, and p is a discount coefficient.4 NCF denotes news about future cash flows (i.e., dividends or consumption),and NR denotes news about future discount rates (i.e., expected returns). This equation, which is an accounting identity rather than a behavioral model, says that unexpected stock returns must be associated with changes in expectationsof futurecash flows or discountrates. An increase in expected future cash flows is associated with a capital gain today, while an increasein discountratesis associated with a capital loss today. The reason is that with a given dividend stream, higher future returnscan be generated only by future price appreciationfrom a lower currentprice. These return components can also be interpreted approximatelyas permanentand transitory shocks to wealth. Returns generated by cash-flow news are never reversed subsequently, whereasreturnsgeneratedby discountrate news are offset by lower returns in the future. From this perspective it should not be surprisingthat conservativelong-terminvestors
4 While Campbell and Shiller (1988a) constrainthe discount coefficient p to values determinedby the average log as dividend yield, p has other possible interpretations well. Campbell (1993, 1996) links p to the average consumptionwealth ratio.In effect, the latterinterpretation be seen as can a slightly modified version of the former.Considera mutual fund that reinvests the dividendspaid by the stocks it holds, and a mutual-fundinvestor who finances her consumption by redeeming a fraction of her mutual-fundshares every year. Effectively, the investor's consumptionis now a dividend paid by the fund, and the investor's wealth (the value of her remainingmutualfund shares)is now the ex-dividend price of the fund. Thus, we can use the loglinear model to describe a portfolio strategy as well as an underlyingasset and let the average consumption-wealthratio generatedby the strategy determine the discount coefficient p, provided that the consumption-wealthratio implied by the strategy does not behave explosively.

VOL.94 NO. 5

CAMPBELL AND VUOLTEENAHO: BAD BETA, GOOD BETA

1253

are more averse to cash-flow risk than to discount-raterisk. To implementthis decomposition,we follow Campbell (1991) and estimate the cash-flownews and discount-rate-news series using a VAR model. This VAR methodology first

(1) to back out the cash-flow news. This practice has an importantadvantage:one does not necessarily have to understand the short-run dynamics of dividends; one need only understand the dynamics of expected returns. We assume that the data are generatedby a first-orderVAR model (2) Zt+l = a + rzt + ut+1

estimates the terms Etrt+ and (Et+ - Et) lJ=1 prt++ and then uses r,+l and equation

where zt+ 1 is a m-by-l state vector with rt+ as its first element, a and r are m-by-l vector and m-by-m matrix of constant parameters, and ut+l an i.i.d. m-by-l vector of shocks. Of course, this formulationalso allows for higherorder VAR models via a simple redefinitionof the state vector to include lagged values. Providedthatthe process in equation(2) generates the data, t + 1 cash-flow and discountrate news are linear functionsof the t + 1 shock vector:
(3)
NCF,t+1 = (el'
NDR,t+

+ el'X)ut+

= el XUt+1.

The VAR shocks are mapped to news by X, definedas k = pr(I - pr)-1. el'X capturesthe long-run significance of each individual VAR shock to discount-rateexpectations.The greater the absolute value of a variable's coefficient in the returnprediction equation (the top row of r), the greaterthe weight the variable receives in the discount-rate-news formula.More persistent variables should also receive more weight, which is capturedby the term (I - pF)-1. B. VARState Variables and Estimation To operationalize the VAR approach, we need to specify the variables to be included in

the state vector. We opt for a parsimonious model with the following four state variables: the excess market return (measured as the log excess return on the Center for Research in Security Prices [CRSP] value-weighted index over Treasurybills); the yield spread between long-term and short-termbonds (measured as the difference between the ten-year constantmaturity taxable bond yield and the yield on short-term taxable notes, in annualized percentage points); the market's smoothed priceearnings ratio (measuredas the log ratio of the S&P 500 price index to a ten-year moving average of S&P 500 earnings);and the small-stock value spread (measured as the difference between the log book-to-marketratios of small value and small growth stocks). The Appendix to this paper, available at www.aeaweb.org/aer, presents full details of data construction.Summary statistics are reportedin Table 1. The three predictor variables can be motivated as follows. First,the yield curve tracksthe business cycle, and there are a number of reasons why expected returnson the stock market could covary with the business cycle. Second, ratioswill necessarilyimply high price-earnings low long-runexpected returns,if expected earnings growth is constant. Third, the small-stock value spread can be motivated by the ICAPM itself. If small growth stocks have low expected returns and small value stocks have high expected returns,and this returndifferentialis not explained by the CAPM betas, the ICAPM requires that the small-growth-stocks'returnpredicts lower and the small-value-stocks' return predicts higher future marketreturns. There are other more direct stories that also suggest the small-stock value spread should be relatedto market-widediscount rates. One possibility is that small growthstocks generatecash flows in the more distant future and therefore their prices are more sensitive to changes in discount rates,just as coupon bonds with a high duration are more sensitive to interest-rate movements than are bonds with a low duration (BradfordCornell, 1999). Anotherpossibility is that small growth companies are particularly dependent on external financing and thus are sensitive to equity marketand broaderfinancial conditions (Victor Ng et al., 1992; Gabriel 2000). A Perez-Quirosand Allan Timmermann,

1254

THEAMERICAN ECONOMICREVIEW
TABLE 1-DESCRIPTIVE STATISTICS THEVAR STATEVARIABLES OF

DECEMBER2004

Variable r-M TY PE VS Correlations

rMt+

Mean 0.004 0.629 2.868 1.653

Median 0.009 0.550 2.852 1.522

rM,t+l

Stdev. 0.056 0.643 0.374 0.374 TYt+ 1 0.071 1 -0.253 0.423 0.065 0.906 -0.263 0.425

Min -0.344 -1.350 1.501 1.192 PE,+ -0.006 -0.253 1 -0.320 0.070 -0.248 0.992 -0.322

Max 0.322 2.720 3.891 2.713

Autocorr. 0.108 0.906 0.992 0.992 VSt + -0.030 0.423 -0.320 1 -0.031 0.420 -0.318 0.992

TYt+ 1 PEt+ 1 V+ 1
rM,t

1 0.071 -0.006 -0.030 0.103 0.070 -0.090 -0.025

TYt PEt VSt

Notes: The table shows the descriptive statistics of the VAR state variables estimated from the full sample period 1928:12-2001:12, 877 monthly data points. rM is the excess log returnon the CRSP value-weight index. TYis the term yield taxablebonds and short-term spreadin percentagepoints, measuredas the yield differencebetween ten-yearconstant-maturity taxable notes. PE is the log ratio of the S&P 500's price to the S&P 500's ten-year moving average of earnings. VS is the small-stock value-spread, the difference in the log book-to-marketratios of small value and small growth stocks. The small-value and small-growthportfolios are two of the six elementaryportfolios constructedby Davis et al. (2000). "Stdev." denotes standarddeviation and "Autocorr." first-orderautocorrelation the series. the of

third possibility is that episodes of irrational investor optimism (Shiller, 2000) have a particularly powerful effect on small growth stocks. Table 2 reports parameterestimates for the VAR model over our full sample periodJanuary 1929 to December 2001. Each row of the table corresponds to a different equation of the model. The first five columns reportcoefficients on the five explanatory variables: a constant; and lags of the excess market return, term yield spread, price-earnings ratio, and smallstock-value spread. Ordinary least squares (OLS) standard errors are reported in square brackets below the coefficients. For comparison, we also report in parenthesesstandarderrors from a bootstrapexercise, details of which are summarizedin the Appendix. Finally, we report the R2 and F statistics for each regression. The bottom of the table reportsthe correlation matrix of the equation residuals, with standard deviations of each residual on the diagonal. The firstrow of Table 2 shows that all four of our VAR state variables have some ability to predict excess returns on the aggregate stock market.Marketreturnsdisplay a modest degree

of momentum; the coefficient on the lagged excess market returnis 0.094, with a standard errorof 0.034. The term yield spreadpositively predicts the market return, consistent with the findings of Donald Keim and Stambaugh (1986), Campbell(1987), and Fama and French (1989). The smoothed price-earningsratio negatively predicts the return, consistent with Campbelland Shiller (1988b, 1998) and related work using the aggregate dividend-price ratio (Michael Rozeff, 1984; Campbell and Shiller, 1988a; Fama and French, 1988, 1989). The small-stockvalue spreadnegatively predictsthe return,consistent with Eleswarapuand Reinganum (2004) and Brennanet al. (2001). Overall, the 2.6-percent R2 of the return forecasting equation is reasonablefor a monthly model. The remaining rows of Table 2 summarize the dynamics of the explanatoryvariables. The term spreadis approximatelyan AR(1) process with an autoregressivecoefficient of 0.88, but the lagged small-stock value spread also has some ability to predict the term spread. The price-earningsratio is highly persistent, with a root very close to unity, but it is also predicted by the lagged marketreturn.This predictability

VOL.94 NO. 5

CAMPBELL AND VUOLTEENAHO: BAD BETA, GOOD BETA TABLE 2-VAR PARAMETER ESTIMATES

1255

Constant
rMt+

e,t 0.094 [0.033] (0.034) 0.046 [0.165] (0.170) 0.519 [0.022] (0.022) -0.005 [0.029] (0.028)
rMt

TY, 0.006 [0.003] (0.003) 0.879 [0.016] (0.017) 0.002 [0.002] (0.002) 0.002 [0.003] (0.003) TY+ 0.018 (0.048) 0.268 (0.013) 0.018 (0.039) -0.012 (0.034)

PE, -0.014 [0.005] (0.007) -0.036 [0.026] (0.031) 0.994 [0.004] (0.004) 0.000 [0.005] (0.006)

VS, -0.013 [0.006] (0.008) 0.082 [0.028] (0.036) -0.003 [0.004] (0.005) 0.991 [0.005] (0.008) PE, 1 0.777 (0.018) 0.018 (0.039) 0.036 (0.002) -0.086 (0.045)

R2 % 2.57

F 5.34

?0.062 [0.020] (0.026) 0.046 [0.097] (0.012) 0.019 [0.013] (0.017) 0.014 [0.017] (0.024)

TYt+ 1

82.41

1.02X 103

PEt+ 1

99.06

2.29x 104

VSt+ 1

98.40

1.34X 104

corr/std i rMt+1 TY+ 1 PEt+ 1 VS+ 1

VSt -0.052 (0.052) -0.012 (0.034) -0.086 (0.045) 0.047 (0.003)

00.055 (0.003) 0.018 (0.048) 0.777 (0.018) -0.052 (0.052)

Notes: The table shows the OLS parameter VAR model including a constant,the log excess market estimates for a first-order return (rM),term yield spread (TY), price-earningsratio (PE), and small-stock value spread (VS). Each set of three rows correspondsto a differentdependentvariable.The first five columns reportcoefficients on the five explanatoryvariables,and the remainingcolumns show R2 and F statistics. OLS standarderrorsare in squarebracketsand bootstrapstandarderrorsin parentheses.Bootstrapstandarderrorsare computedfrom 2,500 simulatedrealizations.The table also reportsthe correlation matrix of the shocks with shock standarddeviations on the diagonal, labeled "corr/std."Sample period for the dependent variables is 1929:1-2001:12, 876 monthly data points.

may reflect short-termmomentum in stock returns, but it may also reflect the fact that the recent history of returnsis correlatedwith earnings news that is not yet reflectedin our lagged earningsmeasure.Finally, the small-stockvalue spreadis also a highly persistentAR(1) process. Table 3 summarizesthe behavior of the implied cash-flow news and discount-rate news componentsof the marketreturn.The top panel shows that discount-ratenews has a standard deviation of about 5 percent per month, much deviation of largerthanthe 2.5-percentstandard cash-flow news. This is consistent with the finding of Campbell(1991) that discount-ratenews is the dominantcomponentof the marketreturn. The table also shows that the two components of returnare almost uncorrelatedwith one another. This finding differs from Campbell

(1991) and particularlyCampbell (1996); it results from our use of a richerforecastingmodel that includes the value spread as well as the price-earningsratio. Table 3 also reportsthe correlationsof each state variable innovation with the estimated news terms, and the coefficients (el' + el'k) and el'k that map innovationsto cash-flow and discount-ratenews. Innovations to returnsand the price-earnings ratio are highly negatively correlated with discount-rate news, reflecting the mean reversion in stock prices that is iminnoplied by our VAR system. Market-return vations are weakly positively correlated with cash-flow news, indicating that some part of a market rise is typically justified by underlying improvements in expected future cash flows. Innovationsto the price-earnings ratio, however,

1256

THEAMERICAN ECONOMICREVIEW
TABLE 3-CASH-FLOW AND DISCOUNT-RATE NEWS FOR THE MARKET PORTFOLIO

DECEMBER2004

News covariance NCF

NDR

NcF 0.00064 (0.00022) 0.00015 (0.00037)

NCF

NDR 0.00015 (0.00037) 0.00267 (0.00070)

NDR

News corr/std
NCF NDR

NcF 0.0252 (0.004) 0.114 (0.232)

NCF

NDR 0.114 (0.232) 0.0517 (0.007)

NDR

Shock correlations rm shock TY shock PE shock VS shock

Functions rm shock TY shock PE shock VS shock

0.352 (0.224) 0.128 (0.134) -0.204 (0.238) -0.493 (0.243)

-0.890 (0.036) 0.042 (0.081) -0.925 (0.039) -0.186 (0.152)

0.602 (0.060) 0.011 (0.013) -0.883 (0.104) -0.283 (0.160)

-0.398 (0.060) 0.011 (0.013) -0.883 (0.104) -0.283 (0.160)

Notes: The table shows the propertiesof cash-flow news (NCF)and discount-ratenews (NDR) implied by the VAR model of Table 2. The upper-left section of the table shows the covariance matrix of the news terms. The upper-rightsection shows the correlationmatrixof the news terms with standard deviationson the diagonal.The lower-left section shows the correlation of shocks to individual state variables with the news terms. The lower-rightsection shows the functions (el' + el'k, el'X) that map the state-variableshocks to cash-flow and discount-ratenews. We define X = pr(I - pr)- , where r is the estimated VAR transitionmatrix from Table 2 and p is set to 0.95 per annum. rm is the excess log returnon the CRSP value-weight index. TYis the term yield spread.PE is the log ratio of the S&P 500's price to the S&P 500's ten-yearmoving average of earnings. VS is the small-stock value-spread,the difference in log book-to-marketsof value and growth stocks. Bootstrap standarderrors (in parentheses)are computed from 2,500 simulatedrealizations.

are weakly negatively correlatedwith cash-flow news, suggesting that price increases relative to earnings are not usually justified by improvements in future earnings growth. Figure 1 illustratesthe VAR model's view of stock markethistory in relationto NBER recessions. Each dottedline in the figurecorresponds to the trough of a recession as defined by the NBER. The top panel reports a trailing exponentially weighted moving average of the market's cash-flow news, while the bottom panel reports the same moving average of the market's discount-rate news. It is clear from the figure that in some recessions our model attributesstock marketdeclines to declining cash flows (e.g., 1991), in others to increasing discount rates (e.g., 2001), and in others to both types of news (e.g., the Great Depression and the 1970s). We might call the first type of recession a "profitability recession,"the second a "valuationrecession," and the third type type a "mixed recession." A valuation recession is characterizedby a declining price-earningsratio, a steepeningyield curve, and largerdeclines in growth stocks than in value stocks. Profit-

ability and valuation recessions, as opposed to mixed recessions, will be particularly influential observations when we estimate cash-flow and discount-ratebetas, because these are episodes in which cash-flow and discount-ratenews do not move closely together. We set p = 0.95 r2 in Table 3 and use the same value throughoutthe paper. Recall that p can be related to either the average dividend yield or the average consumption wealth ratio. An annualizedp of 0.95 correspondsto an avraerage dividend-priceor consumption-wealth tio of 5.2 percent, where wealth is measured after subtracting consumption. We pick the value 0.95 because approximately 5-percent consumption of total wealth per year seems reasonable for a long-term investor, such as a university endowment. II. MeasuringCash-Flowand Discount-Rate Betas We have shown that market returnscontain two components,both of which display substantial volatility and which are not highly corre-

VOL.94 NO. 5

CAMPBELL AND VUOLTEENAHO: BAD BETA, GOOD BETA .I

.

1257

*
,t

0 z
a) 0 0

LI

cn

1940

1950

1960

1970

1980

1990

2000

3-

2? 1

i'

FIGURE 1. CASH-FLOW AND DISCOUNT-RATE RECESSIONS

Notes: This figure plots the cash-flow news and negative of discount-ratenews, smoothed with a trailing exponentially weighted moving average. The decay parameteris set to 0.08 per month, and the smoothed news series are generated as MAt(N)= 0.08Nt + (1 - 0.08)MA,_ i(N). The dotted vertical lines denote NBER business-cycle troughs.The sample period is 1929:1-2001:12.

lated with one another. This raises the possibility that different types of stocks may have differentbetas with the two componentsof the market. In this section we measure cashflow betas and discount-ratebetas separately. We define the cash-flow beta as
(4) Pi,CF-

divides by the total varianceof unexpectedmarket returns,not the variance of cash-flow news or discount-ratenews separately. This implies that the cash-flow beta and the discount-rate beta add up to the total marketbeta (6)
Pi,M
=

r,) Var(ro,t Et-_ rM,t)

Cov(ri,t, NC,t)

Pi,CF

Pi,DR.

and the discount-ratebeta as

(5)

3Pi,DR

(5) -

Cov(r,t, -NDRt) - E-1 re,t) Var(rmt

Note that the discount-ratebeta is defined as the covariance of an asset's return with good news about the stock marketin form of lowerthan-expecteddiscountrates, and that each beta

Our estimates show that there is interesting variation across assets and across time in the two components of the marketbeta. Our main finding is that value stocks have higher cashflow betas than growth stocks. This result is consistent with the empirical results of Randolph Cohen et al. (2003a). Cohen et al. measure cash-flow betas by regressingthe multiyear return on equity (ROE) of value and growth stocks on the market's multiyear ROE. They find that value stocks have higher ROE betas

1258

THEAMERICAN ECONOMICREVIEW

DECEMBER2004

than growth stocks. There is also evidence that value stock returnsare correlatedwith shocks to GDP-growthforecasts (Jimmy Liew and Maria Vassalou, 2000; Vassalou, 2003). This sensitivity of value stocks' cash-flow fundamentalsto economy-wide cash-flow fundamentalsplays a key role in our two-beta model's ability to explain the value premiumin the subsequentpricing tests. We constructtwo sets of portfolios to use as test assets. The first is a standard set of 25 portfolios sorted by marketcapitalization(ME) and book-to-market ratio (BE/ME), available from Professor Kenneth French's Web site (www.mba.tuck.dartmouth.edu/pages/faculty/ ken.french). Kent Daniel and Sheridan Titman (1997) point out that it can be dangerous to test asset pricing models using only portfolios sorted by characteristicsknown to be related to average returns,such as size and value. Characteristicssortedportfolios are likely to show some spread in betas identified as risk by almost any asset pricing model, at least in sample. When the model is estimated, a high premiumper unit of beta will fit the large variation in average returns. Thus, at least when premia are not constrainedby theory, an asset pricing model may spuriously explain the average returns to characteristics-sortedportfolios. To alleviate this concern,we follow the advice of Daniel and Titman (1997) and constructa second set of 20 portfolios sorted on past risk loadings with VAR state variables (excluding the pricesmoothed earnings ratio PE, since highfrequencychanges in PE are so highly collinear with marketreturns).The methodology we use to constructthese portfolios is explained in the Appendix. Both the 25 size- and book-tomarket-sorted returnsand the 20 risk-sortedreturns are measured over the period January 1929 to December 2001. We estimate cash-flow and discount-ratebetas using the fitted values of the market's cashflow and discount-ratenews. Specifically, we use sample covariances and variances in the formulas (4) and (5), allowing for one additional lag of the news terms. The additionallag is motivated by the possibility that, especially duringthe early years of our sample period, not all stocks in our test-assetportfolioswere traded

frequently and synchronously.5 Full details of the beta construction are reported in the Appendix. When we apply this estimation technique to our test-asset returnsand our estimated market news series, we find cash-flow and discount-rate dramatic differences in the beta estimates between the first half of our 1929-2001 sample and the second half. Accordingly, we report betas separately for two subsamples, January 1929-June 1963 and July 1963-December 2001. We choose to split the sample at July 1963, because that is when COMPUSTATdata become reliable and because most of the evidence on the book-to-marketanomaly is obtained from the post-1963 period. Unlike the thoroughly mined second subsample, the first subsample is relatively untouchedand presents an opportunityfor an out-of-sample test. The top half of Table 4 shows the estimated betas for the 25 size- and book-to-marketportfolios over the 1929-1963 period. The portfolios are organized in a square matrix with growth stocks at the left, value stocks at the right, small stocks at the top, and large stocks at the bottom. At the right edge of the matrix we report the differences between the extreme growth and extreme value portfolios in each size group; along the bottom of the matrix we report the differences between the extreme small and extreme large portfolios in each BE/ME category. The top matrixdisplays cashflow betas, while the bottom matrix displays discount-rate betas. In square brackets after each beta estimate we report a standarderror, calculatedconditionalon the realizationsof the news series from the aggregateVAR model. In the pre-1963 sample period, value stocks
5 If some portfolio returns are contaminated by stale prices, marketreturnand news termsmay spuriouslyappear to lead the portfolioreturns,as noted by Myron Scholes and Joseph Williams (1977) and Elroy Dimson (1979). In addition, Andrew Lo and A. Craig MacKinlay(1990) show that the transactionprices of individual stocks tend to react in partto movements in the overall marketwith a lag, and the smaller the company, the greater is the lagged price reaction. GrantMcQueen et al. (1996) and James Peterson and Gary Sanger (1995) show that these effects exist even in relatively low-frequencydata (i.e., those sampledmonthly). These problems are alleviated by the inclusion of the lag term.

VOL.94 NO. 5

CAMPBELL AND VUOLTEENAHO: BAD BETA, GOOD BETA

TABLE4-CASH-FLOW AND DISCOUNT-RATE BETASIN THEEARLYSAMPLE

1259

cCF

Growth 0.53 0.30 0.30 0.20 0.20 -0.33 [0.11] [0.06] [0.06] [0.05] [0.05] [0.09] 0.46 0.34 0.28 0.26 0.19 -0.26

2 [0.09] [0.06] [0.05] [0.05] [0.05] [0.06]

2

3 0.40 0.36 0.31 0.31 0.28 -0.12 3 [0.19] [0.11] [0.08] [0.08] [0.07] [0.15] 1.32 1.09 1.08 0.97 0.87 -0.43 [0.15] [0.11] [0.09] [0.09] [0.08] [0.10] 3 [0.05] [0.04] [0.05] [0.04] 2 0.31 0.25 0.26 0.27 3 [0.07] [0.07] [0.07] [0.06] 1.04 0.96 1.00 0.88 [0.09] [0.08] [0.09] [0.08] 1.20 1.09 1.17 1.09 [0.06] [0.06] [0.06] [0.05] 0.37 0.28 0.31 0.32 1.27 1.25 1.05 1.06 1.06 -0.21 [0.08] [0.06] [0.06] [0.05] [0.06] [0.05] 0.42 0.38 0.35 0.35 0.33 -0.09
4

4 [0.07] [0.06] [0.06] [0.07] [0.07] [0.04]

Value 0.49 0.45 0.47 0.50 0.40 -0.10

Value

Diff. -0.04 0.16 0.18 0.30 0.19 [0.07] [0.04] [0.04] [0.05] [0.06]

Small 2 3 4 Large Diff.

IDR

[0.08] [0.08] [0.08] [0.09] [0.09] [0.04]

Growth

Diff.

Small 2 3 4 Large Diff.

1.32 1.04 1.13 0.87 0.88 -0.45

[0.18] [0.11] [0.10] [0.07] [0.07] [0.15]

1.46 1.15 1.01 0.97 0.82 -0.64

[0.14] [0.11] [0.10] [0.10] [0.09] [0.08] 4 [0.07] [0.06] [0.07] [0.06] 4 [0.11] [0.09] [0.10] [0.10]

1.27 1.25 1.27 1.36 1.18 -0.08

[0.15] [0.13] [0.12] [0.13] [0.12] [0.10]

-0.06 0.21 0.14 0.49 0.31

[0.15] [0.08] [0.06] [0.10] [0.10]

/CF

Lo rM 0.21 0.15 0.18 0.16 [0.04] [0.03] [0.04] [0.04] Lo br 0.73 0.64 0.73 0.65 [0.06] [0.05] [0.06] [0.06] 0.87 0.75 0.85 0.76 0.25 0.19 0.21 0.21

Hi r 0.45 0.37 0.41 0.40

Hi

Diff. 0.25 0.22 0.23 0.24 [0.05] [0.05] [0.04] [0.05] Diff. 0.73 0.66 0.64 0.69 [0.09] [0.08] [0.08] [0.09]

Lo bs Hi bvs Lo br Hi bTy
JDR

[0.09] [0.08] [0.08] [0.08] rM [0.13] [0.11] [0.12] [0.12]

Lo bs Hi bvs Lo bTy Hi bTy

1.46 1.30 1.38 1.34

Notes: The table shows the estimated cash-flow (crF) and discount-ratebetas (IDR) for the 25 ME- and BE/ME-sorted denotes the lowest BE/ME, "value"the highest BE/ME, "small"the lowest portfolios and 20 risk-sortedportfolios. "Growth" on are ME, and "large"the highest ME stocks. bvs, by, and brM past return-loadings value-spreadshock, term-yield shock, and market-return shock. "Diff."is the difference between the extreme cells. Standarderrors[in brackets]are conditionalon the estimatednews series. Estimates are for the 1929:1-1963:6 period.

have both higher cash-flow and higherdiscountrate betas than growth stocks. An equalweighted average of the extreme value stocks across size quintiles has a cash-flow beta 0.16 higher than an equal-weighted average of the extreme growth stocks. The difference in estimated discount-ratebetas is 0.22 in the same direction. Similar to value stocks, small stocks have higher cash-flow betas and discount-rate betas than large stocks in this sample (by 0.18 and 0.36, respectively, for an equal-weighted average of the smallest stocks across value quintiles relative to an equal-weighted average of the largest stocks). In summary, value and small stocks were unambiguously riskier than

growth and large stocks over the 1929-1963 period. A partialexception to this statementinvolves the smallest growth portfolio, which is particularly risky and has both cash-flow and discountratebetas thatexceed those of the smallest value portfolio. This small growth portfolio is well known to present a particular challenge to asset pricing models, for example, the three-factor model of Fama and French (1993), which does not fit this portfolio well. Recent evidence on small growthstocks by Owen Lamontand Richard Thaler (2003), Mark Mitchell et al. (2002), Eugene D'Avolio (2002), and others suggests that the pricing of some small growth stocks is

1260

THEAMERICAN ECONOMICREVIEW
TABLE 5-CASH-FLOW AND DISCOUNT-RATE BETAS IN THE MODERN SAMPLE

DECEMBER2004

(CF

Growth 0.06 0.04 0.03 0.03 0.03 -0.03 [0.07] [0.06] [0.05] [0.05] [0.04] [0.05] 0.07 0.08 0.09 0.10 0.08 0.02

2 [0.06] [0.05] [0.04] [0.04] [0.03] [0.05]

2

3 0.09 0.10 0.11 0.11 0.09 -0.01 3 [0.11] [0.09] [0.08] [0.08] [0.07] [0.11] 1.18 1.07 0.95 0.89 0.74 -0.44 3 [0.03] [0.03] [0.03] [0.03] 0.10 0.07 0.08 0.09 3 [0.06] [0.07] [0.07] [0.06] 0.88 1.06 1.05 0.91 [0.07] [0.07] [0.07] [0.06] 1.12 1.30 1.23 1.11 [0.04] [0.04] [0.04] [0.03] 0.10 0.05 0.08 0.08 [0.10] [0.08] [0.08] [0.07] [0.06] [0.10] 1.12 0.96 0.82 0.79 0.63 -0.49 [0.05] [0.04] [0.04] [0.03] [0.03] [0.04] 0.09 0.11 0.12 0.11 0.11 0.02

4 [0.04] [0.04] [0.03] [0.03] [0.03] [0.04]

4

Value 0.13 0.12 0.13 0.13 0.11 -0.01

Value

Diff. 0.07 0.09 0.09 0.10 0.09 [0.04] [0.03] [0.04] [0.04] [0.03]

Small 2 3 4 Large Diff.

&DR

[0.04] [0.04] [0.04] [0.04] [0.03] [0.04]

Growth

Diff.

Small 2 3 4 Large Diff.

kCF

1.66 1.54 1.41 1.27 1.00 -0.66

[0.13] [0.11] [0.10] [0.09] [0.07] [0.12]

1.37 1.22 1.11 1.05 0.87 -0.50

2

[0.09] [0.08] [0.07] [0.07] [0.07] [0.09]

4

1.12 1.03 0.94 0.87 0.68 -0.44

Hibr

[0.10] [0.09] [0.09] [0.08] [0.07] [0.08]

-0.54 -0.52 -0.47 -0.41 -0.33

[0.08] [0.08] [0.09] [0.09] [0.08]

Lo br

Diff.

Lo bvs Hi bvs Lo by Hi bTy

[DR

0.09 0.06 0.06 0.09

[0.03] [0.03] [0.03] [0.03]

0.08 0.06 0.04 0.07 2 0.77 0.85 0.86 0.79

[0.04] [0.05] [0.04] [0.04] 4 [0.08] [0.09] [0.08] [0.07]

0.12 0.06 0.06 0.10

Hibr

[0.05] [0.06] [0.06] [0.05]

0.04 -0.01 0.00 0.00

[0.04] [0.04] [0.04] [0.04] Diff.

Lo bvs Hi bvs Lo bY Hi bTy

LobrM 0.57 [0.06] 0.67 [0.06] 0.73 0.61 [0.07] [0.06]

1.40 1.58 1.60 1.39

[0.09] [0.11] [0.12] [0.09]

0.82 0.91 0.87 0.78

[0.08] [0.11] [0.10] [0.08]

Notes: The table shows the estimated cash-flow (/CF) and discount-ratebetas (I%R) for the 25 ME- and BE/ME-sorted denotes the lowest BE/ME, "value"the highest BE/ME, "small"the lowest portfolios and 20 risk-sortedportfolios. "Growth" on ME, and "large"the highest ME stocks. bvs, bTy, and br are past return-loadings value-spreadshock, term-yield shock, and market-return shock. "Diff." is the difference between the extreme cells. Standarderrors[in brackets]are conditionalon the estimated news series. Estimates are for the 1963:7-2001:12 period.

materiallyaffected by short-saleconstraintsand other limits to arbitrage.This may help to explain the unusual behavior of the small growth portfolio. The bottom half of Table 4 shows the cashflow and discount-ratebetas for the risk-sorted portfolios. Both cash-flow betas and discountrate betas are high for stocks that have had high market betas in the past. Thus, in the early sample period, sorting stocks by their past market betas induces a spread in both cash-flow betas and discount-ratebetas. Sorting stocks by their value-spreador term-spreadsensitivity induces only a relatively modest spread in either beta.

The patterns are completely different in the post-1963 period shown in Table 5. In this subsample, value stocks still have slightly higher cash-flow betas than growth stocks, but much lower discount-rate betas. The difference in cash-flow betas between the average across extreme-value portfolios and the average across extreme-growth portfolios is a modest 0.09. What is remarkable is that the pattern of discount-rate betas reverses in the modem period, so that growth stocks have significantly higher discount-rate betas than value stocks. The difference is economically large (0.45) and statistically significant. Recall that cash-flow and discount-rate betas sum up to the CAPM

VOL.94 NO. 5

CAMPBELL AND VUOLTEENAHO: BAD BETA, GOOD BETA

1261

beta; thus growth stocks have higher market betas in the modem period, but their betas are disproportionatelyof the "good" discount-rate variety ratherthan the "bad"cash-flow variety. The changes in the risk characteristics of value and growth stocks that we identify by comparingthe periodsbefore and after 1963 are consistent with recent research by Francesco Franzoni (2004). Franzoni points out that the market betas of value stocks and small stocks have declined over time relative to the market betas of growth stocks and large stocks. We extend his research by exploring time changes in the two componentsof marketbeta: the cashflow beta and the discount-ratebeta. What economic forces have caused these changes in betas? We suspect that the changing characteristicsof value and growth stocks and small and large stocks are related to these patterns in sensitivities. Our first subsample is dominatedby the Great Depression and its aftermath.Perhapsin the 1930s value stocks were fallen angels with a large debt load accumulated during the Great Depression. The higher leverage of value stocks relative to that of growth stocks could explain both the higher cash-flow and expected-return betas of value stocks from 1929 to 1963. In general, low leverage and strong overall position of a company may lead to a low cash-flow beta, and high leverage and weak position to a high cash-flow beta. We also hypothesize that future investment opportunities,long durationof cash flows, and dependenceon externalequity finance lead to a high discount-ratebeta. For example, if a distressed firm needed new equity financing simply to survive afterthe GreatDepression, and if the availability and cost of such financingwere relatedto the overall cost of capital, then such a firm's value was likely to have been very sensitive to discount-rate news. Similarly, new small firms with a negative current cash flow but valuable investmentopportunitiesare likely to be very sensitive to discount-rate news. In the moder subsample,the growth portfolio probably contains a higher proportionof young companies following the initial-public-offering (IPO) wave of the 1960s, the inclusion of NASDAQ firms in our sample during the late 1970s, and the flood of technology IPOs in the 1990s.

The increase in growth stocks' discount-rate betas may also be partially explained by changes in stock market listing requirements. During the early period, only firms with significant internalcash flow made it to the Big Board and thus our sample. This is because, in the past, the New York Stock Exchange (NYSE) had very strict profitabilityrequirementsfor a firm to be listed on the exchange. The lowBE/ME stocks in the first half of the sample are thus likely to be consistently profitableand independentof externalfinancing.In contrast,our post-1963 sample also contains NASDAQ stocks and less-profitable new lists on the NYSE. These firms are listed precisely to improve their access to equity financing,and many of them will not even survive-let alone achieve their growth expectations-without a continuing availability of inexpensive equity financing. Finally, it is possible that our discount-rate news is simply news about investor sentiment. If growth investing has become more popular among irrational investors during our sample period, growth stocks may have become more sensitive to shifts in the sentiment of these investors. Our risk-sortedportfolios also have different betas in the second subsample.Sorting on market risk while controlling for other state variables induces a spreadin only the discount-rate beta in the second subsample.
III. Pricing Cash-Flow and Discount-Rate Betas

We have shown that in the period since 1963, there is a strikingdifference in the beta composition of value and growth stocks. The market betas of growth stocks are disproportionately composed of discount-rate betas rather than cash-flow betas. The opposite is true for value stocks. Motivated by this finding, we next examine the validity of a long-horizon investor's firstorder condition, assuming that the investor holds a 100-percent allocation to the market portfolio of stocks at all times. We ask whether the investor would be better off adding a margin-financedposition in some of our test assets (such as value or small stocks), as a

1262

THEAMERICAN ECONOMICREVIEW

DECEMBER2004

short-horizon investor's first-order condition would suggest. Our main finding is that the long-horizon investor's first-ordercondition is not violated by our test assets and that the difference in beta compositioncan largely explain the high returns on value and low returnson growth stocks relative to the predictionsof the static CAPM. The extreme small-growthportfolio remains an outlier even in our model, but the returnson this portfolio are not sufficientlyanomalousto cause a statisticalrejection of the model. A. An IntertemporalAsset Pricing Model Campbell (1993) derives an approximate discrete-time version of Merton's (1973) ICAPM. The model's central pricing statement is based on the first-ordercondition for an investor who holds a portfoliop of tradableassets that contains all of her wealth. Campbell assumes that this portfolio is observable in order to derive testable asset-pricing implications from the first-ordercondition. Campbell considers an infinitely lived investor who has the recursive preferencesproposed by Epstein and Zin (1989, 1991), with time discount factor 6, relative risk aversion y, and substitution i. Campelasticity of intertemporal bell assumes that all asset returnsare conditionally lognormal,and that the investor's portfolio returns and its two components are homoskedastic. The assumptionof lognormality can be relaxed if one is willing to use Taylor approximations to the true Euler equations, and the model can be extended to allow changing variances as discussed by Chen (2003). Empirically, changes in volatility seem to be much less persistentthan changes in expected returns,and thus they generate relatively modest intertemporal hedging effects on portfolio demands (George Chacko and Viceira, 1999). For this reason we continue to assume constant variances in the empirical work of this paper. Campbellderives an approximatesolution in which risk premia depend only on the coefficient of relativerisk aversion y and the discount coefficient p, and not directlyon the elasticity of substitutionq. The approximation intertemporal is accurateif the elasticity of intertemporal sub-

stitutionis close to one, and it holds exactly in the limit of continuoustime (MarkSchroderand Costis Skiadas, 1999) if the elasticity equals one. In the f = 1 case, p = 8 and the optimal consumption-wealthratio is conveniently constantand equal to 1 - p. Thus our choice of p =
0.951/12

implies that at the end of each month,

the investorchooses to consume 0.43 percentof her wealth if ti = 1. Under these assumptions, the optimality of portfolio strategyp requires that the risk premium on any asset i satisfies

(7)

E,[ri,t+] - rf,t+1+ 2

o.,t

= yCov,(ri,t+ 1, rp,t+ - Etrp,t+ 1)

+ (1 - y)Covt(ri,+ 1, -Np,DR,t+ 1)

where p is the optimal portfolio that the agent = (Et+ - Et) chooses to hold and Np,R,t+I returnnews on this portfolio. The left-hand side of (7) is the expected excess log returnon asset i over the riskless interest rate, plus one-half the variance of the excess returnto adjust for Jensen's inequality. This is the appropriate measure of the risk premium in a lognormal model. The right-hand side of (7) is a weighted average of two covariances: the covarianceof returni with the return on portfoliop, which gets a weight of y, and the covarianceof returni with negative news about future expected returns on portfolio p, which gets a weight of (1 - y). These two covariances represent the myopic and intertemporalhedging components of asset demand, respectively. When y = 1, it is well known that portfolio choice is myopic and the first-ordercondition collapses to the familiar one used to derive the pricing implications of the CAPM. We can rewriteequation (7) to relate the risk premiumto betas with cash-flow news and discovariance with news covariances, and then multiplying and dividing by the conditional variance of portfolio p's return, o2pt, we have
J 1 prp, + I+j is discount-rate or expected-

count-rate news. Using rp,t+1 - Etrp,t+ = Np,cF,t+ - Np,DR,t+I to replace the portfolio

VOL.94 NO. 5

CAMPBELL AND VUOLTEENAHO: BAD BETA, GOOD BETA

1263

(8)

E,[ri,t+] -rft
= y7ap,i,CFp,,

+1

2
Op,tPi,DRp,t

Equation (8) delivers our prediction that "bad beta" with cash-flow news should have a risk price y times greaterthanthe risk price of "good beta" with discount-rate news, which should equal the variance of the returnon portfolio p. In our empiricalwork, we begin by assuming that portfolio p is fully invested in a valueweighted equity index. This assumptionimplies that the risk price of discount-ratenews should equal the varianceof the value-weightedindex, about 5 percent in the early subsample and 2.5 percentin the modem subsample.The only free parameterin equation (8) is then the coefficient of relative risk aversion, y. An alternative assumption would be that portfolio p places a weight w on the valueweighted index and (1 - w) on Treasurybills. If the real Treasury-bill return is constant, this would imply that the variance of portfolio p is w2 times the varianceof the index return,while the cash-flow and discount-rate betas of test asset i with portfolio p are (1/w) times the cash-flow and discount-rate betas with the index return.Under this alternativethe risk prices for both cash-flow and discount-ratebetas are w times smaller, but the risk price for the cashflow beta is still y times the risk price for the discount-ratebeta. The risk prices of the two betas can be used to identify the two free parametersw and y. B. Empirical Estimates of Risk Premia Would an all-stock investor be better off holding stocks at marketweights or overweighting value and small stocks? We examine the validity of an unconditionalversion of the firstorder condition (8) relative to the marketportfolio of stocks. We modify equation(8) in three ways. First, we use simple expected returns, Et[Ri,t+1 Rf,t+], on the left-hand side, inIn the lognormal model, both expectations are the same, and by using simple returnswe make our results easier to comparewith previous emstead of log returns, Et[ri,t+ ] - rft+ 1 + O,t/2.

pirical studies. Second, we condition down equation (7) and derive an unconditional version of equation (8) to avoid estimation of all required conditional moments. Finally, we change the subscriptp to M and use all-stock investment in the marketportfolio of stocks as the reference portfolio, reflecting the fact that we test the optimalityof the marketportfolio of stocks for the long-horizon investor. These modificationsyield (9) E[Ri - Rf] =
Y72Mi,CFM +
-'Mi,DRM.

We assume that the log real risk-free rate is approximatelyconstant.We make this assumption mainly because monthly inflation data are unreliable, especially over our long December 1928 to December 2001 sample period. As shown in the Appendix, this assumption has very little effect on our resultsbecause we focus on stock portfolios. The main practicalimplication of the constant-real-rate assumptionis that cash-flow and discount-rate news computed from excess CRSP value-weight index returns are identically equivalent to news terms computed from real CRSP value-weight index returns. We use 45 test assets, 25 size- and book-tomarket-sorted portfolios and 20 risk-sorted portfolios on the left-hand side of the uncondicondition (9). We evaluate the tional first-order performanceof the traditionalCAPM that restrictscash-flow and discount-ratebetas to have the same price of risk; our two-beta intertemporal asset pricing model that restrictsthe price risk to equal the varianceof the of discount-rate market return; and an unrestricted two-beta model that allows free risk prices for cash-flow betas. As discussed above, the and discount-rate as model can be interpreted a slight unrestricted generalizationof our model that allows the rational investor's portfolio to include Treasury bills as well as equities. Each model is estimated in two different forms: one with a restrictedzero-betarate equal to the Treasury-billrate, and one with an unrestricted zero-beta rate following Fischer Black (1972). The first specificationincludes Treasury bills in the set of alternativeassets available to the investor, while the second assumes that the

1264

THEAMERICAN ECONOMICREVIEW TABLE 6-ASSET PRICING FOR EARLY SAMPLE TESTS THE

DECEMBER2004

Parameter
Rzb

Factor model 0.0042 4.98% [0.0032] (0.0029) 0.0173 20.76% [0.0231] (0.0266) -0.0003 -0.41% [0.0092] (0.0088) 48.08% 0.0117 [0.019] (0.019) 0 0% N/A N/A 0.0069 8.22% [0.221] (0.0248) 0.0066 7.93% [0.0067] (0.0071) 40.26% 0.0126 [0.024] (0.024)

Two-beta ICAPM 0.0023 2.76% [0.0024] (0.0030) 0.0083 9.91% [0.0167] (0.0221) 0.0041 4.95% [0.0006] (0.0006) 45.85% 0.0119 [0.024] (0.031) 0 0% N/A N/A 0.0148 17.80% [0.0175] (0.0442) 0.0041 4.95% [0.0006] (0.0006) 37.98% 0.0133 [0.033] (0.099) 0.0023 2.74% [0.0028] (0.0028) 0.0051 6.11% [0.0046] (0.0046) 0.0051 6.11% [0.0046] (0.0046) 44.52% 0.0127 [0.021] (0.021)

CAPM 0 0% N/A N/A 0.0067 8.00% [0.0034] (0.0034) 0.0067 8.00% [0.0034] (0.0034) 40.26% 0.0126 [0.027] (0.027)

less Rrf(go) % per annum Std. err. A Std. err. B

fcF premium(gl) % per annum Std. err. A Std. err. B
/DR premium(g2) % per annum Std. err. A Std. err. B

R~2
Pricing error 5% critic. val. A 5% critic. val. B

Notes: The table shows premia estimated from the 1929:1-1963:6 sample for an unrestrictedfactor model, the two-beta ICAPM, andthe CAPM. The test assets are the 25 ME- andBE/ME-sorted portfoliosand 20 risk-sortedportfolios.The second column per model constrainsthe zero-beta rate (Rzb)to equal the risk-free rate (R,f). Estimates are from a cross-sectional regressionof average simple excess test-assetreturns(monthlyin fractions)on an interceptand estimatedcash-flow (^cF) and discount-ratebetas (IDR). Standarderrors and critical values [A] are conditional on the estimated news series and (B) incorporatingfull estimation uncertaintyof the news terms. The test rejects if the pricing error is higher than the listed 5 percent critical value.

investor is consideringonly reallocationsof the portfolio among alternative types of equities. Thus in the first specificationwe ask the model to explain the unconditionalequity premiumas well as the premiato value stocks, small stocks, and risk-sortedstocks; in the second specification we remove the equity premiumfrom the set of phenomenato be explained. Table 6 reports results for the early sample period 1929-1963. The table has 6 columns, 2 specifications for each of our 3 asset pricing models. The first 12 rows of Table 6 are divided into 3 sets of 4 rows. The first set of 4 rows correspondsto the zero-beta rate (in excess of the Treasury-bill rate), the second set to the premiumon cash-flow beta, and the third set to the premium on discount-ratebeta. With each set, the first row reports the point estimate in fractions per month, and the second row annualizes this, multiplying by 1,200 to ease the interpretationof the estimate. The third and fourth rows present two alternative standard errorsof the monthly estimate.

These parameters estimatedfrom a crossare sectional regression

(10)

Ri

= go + 9gli,CF

+ g2[i,DR

+ ei

where bar denotes time-series mean and Ri Ri - R,r denotes the sample average simple excess returnon asset i. The implied risk-aversion coefficient can be recovered as g/g2. Standard errorsare producedwith a bootstrap from 2,500 simulated realizations. Our bootstrapexperimentsamples test-asset returnsand VAR errors,and uses the OLS VAR estimates in Table 2 to generate the state-variabledata. We partitionthe VAR errors and test-asset returnsinto two groups, one for 1929 to 1963 and anotherfor 1963 to 2001, which enables us to use the same simulatedrealizationsin subperiod analyses. The first set of standarderrors (labelled A) conditions on estimated news terms and generates betas and return premia separately for each simulated realization, while the second set (labelled B) also estimates the VAR

VOL.94 NO. 5

CAMPBELL AND VUOLTEENAHO: BAD BETA, GOOD BETA

1265

and the news terms separately for each simulated realization. Standarderrors B thus incorporate the considerable additional sampling uncertaintydue to the fact that the news terms as well as betas are generatedregressors. Below the premiaestimates, we reportthe R2 statistic for a cross-sectional regression of average returnson our test assets onto the fitted values from the model. The regression R2 is computed as (11) R2 = 1 -

(Re
i

'2

) (Re ) __
i

which allows for negative R2 for poorly fitting models estimated under the constraintthat the zero-beta rate equals the risk-free rate. Although the regression R2 is intuitive and it transparent, gives equal weight to each asset included in the set of test assets, even though some assets may be more volatile than others. To address this concern we also report a composite pricing error and its 5-percent critical value. The composite pricing erroris computed as e' -le, where e is the vector of estimated residuals from regression (10) and hf is a diagonal matrixwith estimatedreturnvolatilities on the main diagonal. The weighting matrix, ~- 1, in the composite pricing error formula places less weight on noisy observations, yet it is independent of the specific pricing model. We avoid using a freely estimated variance-covariance matrix of test asset returns for Al because with 45 test assets, we are concerned that the inverse of this matrix would be poorly behaved. Robert Hodrick and Xiaoyan Zhang (2001) discuss related alternative methods for assessing the performance of asset pricing models. Two alternative5-percent critical values for the composite pricing errorare producedwith a bootstrap method similar to the one we have described above, except that the test-asset returns are adjusted to be consistent with the pricing model before the random samples are generated. Critical values A condition on estimated news terms, while critical values B take account of the fact that news terms must be estimated.

Table 6 shows that in the 1929-1963 period, the traditionalCAPM explains the cross-section of stock returnsreasonablywell, and is comparable to the restricted two-beta model and the two-beta model with unrestrictedrisk prices. The cross-sectional R2 statistics are about 40 percentfor models with zero-betarates equal to the Treasury-bill rate, and around45 percentfor models with unrestrictedzero-beta rates. None of the models in the table comes close to being rejected at the 5-percent level. Figure 2 provides a visual summaryof these results. The figure plots the predicted average excess return on the horizontal axis and the actual sample average excess returnon the vertical axis. For a model with a 100-percentestimated R2, all the points would fall on the 45degree line displayed in each graph. The triangles in the figures denote the 24 FamaFrench portfolios, and asterisks denote the 20 risk-sortedportfolios. All the models generate nearly identical scatterplots. The good performanceof the CAPM in the 1929-1963 period is due to the fact that in this period, the bad cash-flow beta is roughly a constant fraction of the CAPM beta across assets. Thus our tests cannotdiscriminatebetween the static and intertemporalCAPM models in this period. Results are very different in the 1963-2001 period. Table 7 shows that in this period, the CAPM fails disastrouslyto explain the returns on the test assets. When the zero-betarate is left a free parameter,the cross-sectional regression picks a negative premium for the CAPM beta and implies a near-zeroestimatedR2. When the zero-betarateis constrainedto the risk-freerate, the CAPM R2 falls to -60 percent, i.e., the model has a larger pricing error than the null hypothesis that all portfolios have equal expected returns. The static CAPM is easily rejected at the 5-percent level by both sets of critical values. The two-beta model with a restricted risk price for discount-ratenews explains almost 50 percent of the cross-sectional variationin average returns across our test assets. The model performs almost as well with a restrictedzerobeta rate, equal to the Treasury-billrate, as it does with an unrestricted Treasury-bill rate. This indicatesthatboth the unconditionalequity

1266

THEAMERICAN ECONOMICREVIEW

DECEMBER2004

15 10 5 withzero-beta rate ICAPM

15 10 5 with risk-freerate ICAPM

15 10 5 CAPMwithzero-beta rate

15 10 5 CAPMwith risk-freerate

FIGURE PERFORMANCE THECAPM AND ICAPM, 1929:1-1963:6 2. OF

Notes: The four diagramscorrespondto (clockwise from the top left) the ICAPM with a free zero-betarate, the ICAPM with the zero-beta rate constrainedto the risk-free rate, the CAPM with a constrainedzero-beta rate, and the CAPM with an unconstrainedzero-betarate. The horizontalaxes correspondto the predictedaverage excess returnsand the vertical axes to the sample average realized excess returns.The predictedvalues are from regressionspresentedin Table 6. Trianglesdenote the 25 ME and BE/ME portfolios and asterisks the 20 risk-sortedportfolios.

premium and the premia on alternativeequity portfolios can be rationalizedby the same coefficient of risk aversion. The estimated risk price for cash-flow beta is high at 58 percentper year with a restricted zero-beta rate and 69 percent per year with an unrestrictedzero-beta rate. There are large standarderrors on these estimates, but they are statistically distinguishable from the low risk price on discount-rate news. The model is not rejectedat the 5-percent level by either set of critical values. The criticalvalues for the restrictedintertemporal model with a restrictedzero-beta rate are particularlylarge, an order of magnitudelarger thanthose for the othermodels in the table. This is due to the fact that this model pins down both

the zero-beta rate and the risk price for discount-ratenews, and thus it pins down the total returngeneratedby a unit of discount-ratebeta. Since estimated discount-ratebetas are noisy, estimates of this model can behave extremely badly even if the model is true. risk The two-beta model with an unrestricted an even lower risk price to disprice assigns count-ratebeta than the variance of the market return.This would be consistentwith a modified model in which a conservativerationalinvestor holds a portfolio that contains Treasurybills as well as equities. The implied shareof equities in the portfolio is 60 percent in the model with a restrictedzero-beta rate, and slightly below 40 percent in the model with an unrestrictedzero-

VOL.94 NO. 5

CAMPBELL AND VUOLTEENAHO: BAD BETA, GOOD BETA TABLE 7-ASSET PRICING TESTS THE SAMPLE FOR MODERN

1267

Parameter
R,b less Rf (go) % per annum Std. err. A Std. err. B
ICF premium (gl) % per annum Std. err. A Std. err. B

Factor model 0.0009 1.05% [0.0029] (0.0033) 0.0529 63.47% [0.0178] (0.0325) 0.0007 0.88% [0.0033] (0.0085) 52.10% 0.0271 [0.028] (0.030) 0 0% N/A N/A 0.0572 68.59% [0.0163] (0.0444) 0.0012 1.44% [0.0031] (0.0099) 51.59% 0.0269 [0.042] (0.071)

Two-beta ICAPM -0.0009 -1.04% [0.0031] (0.0031) 0.0575 69.04% [0.0182] (0.0262) 0.0020 2.43% [0.0002] (0.0002) 49.26% 0.0272 [0.051] (0.051) 0 0% N/A N/A 0.0483 57.92% [0.0272] (0.0423) 0.0020 2.43% [0.0002] (0.0002) 47.41% 0.0275 [0.314] (0.488) 0.0069 8.24% [0.0026] (0.0026) -0.0007 -0.83% [0.0034] (0.0034) -0.0007 -0.83% [0.0034] (0.0034) 3.10% 0.0592 [0.032] (0.032)

CAPM 0 0% N/A N/A 0.0051 6.10% [0.0023] (0.0023) 0.0051 6.10% [0.0023] (0.0023) -61.57% 0.0875 [0.046] (0.046)

premium (g2) % per annum Std. err. A Std. err. B

/DR f?2

Pricing error 5% critic. val. A 5% critic. val. B

Notes: The table shows premia estimated from the 1963:7-2001:12 sample for an unrestrictedfactor model, the two-beta ICAPM,and the CAPM. The test assets are the 25 ME- and BE/ME-sorted portfoliosand 20 risk-sortedportfolios.The second column per model constrains the zero-beta rate (RZb) equal the risk-free rate (R,f). Estimates are from a cross-sectional to regressionof average simple excess test-assetreturns(monthlyin fractions)on an interceptand estimatedcash-flow (crF) and discount-ratebetas (/DR). Standarderrors and critical values [A] are conditional on the estimated news series and (B) incorporatingfull estimation uncertaintyof the news terms. The test rejects if the pricing error is higher than the listed 5-percent critical value.

beta rate. This model generates cross-sectional R2 statistics slightly above 50 percent. A visual summary these resultsis providedby of Figure3. Another way to evaluate the performanceof our model is to compare it to less theoretically structuredmodels. We do this in two ways. First, we compare our restrictedICAPM model to a model whose factors are the four innovations from our VAR system, with unrestricted risk prices. In the modem sample, the four unrestricted risk prices line up almost perfectly with those implied by our restrictedmodel. Second, we compare the two-beta model to the influential three-factor model of Fama and French (1993). In the early subsample, the cross-sectionalR2 statistic for the Fama-French three-factor model is 10 percentage points higher than that for our two-beta model with an unconstrainedzero-beta rate, and 1 percentage point higher with a zero-betarate constrainedto the risk-freerate. In the modem subsample,the Fama-Frenchmodel outperformsthe two-beta model by 30 and 26 percentagepoints, respec-

tively. This difference in explanatorypower is not statisticallysignificant,as the restrictionsof our model are not rejected by our composite pricing error test. Given that the Fama-French model has three freely estimatedbetas and thus two additionaldegrees of freedom, we consider the relative performance of the two-beta ICAPM to be a success. Although the two-beta model is generally quite successful in explaining the cross-section of average returns,the model cannot price the extreme small-growthportfolio. In the first subsample, the extreme small-growthportfolio has an annualizedaveragereturnthat is 8.8 percentage points lower than the model's prediction.In the second subsample, the returnon this portfolio is 3.2 percentage points lower than the model's prediction.These pricing-error calculations use the model specificationwith the zerobeta rate constrained to the risk-free rate. In both subsamples, these pricing errors are economically large and not meaningfully smaller than the pricing errors of the Sharpe-Lintner CAPM for this portfolio (9.9 percentagepoints

1268

THEAMERICAN ECONOMICREVIEW A

DECEMBER2004

10 5 ICAPM withzero-beta rate A

5 10 ICAPM with risk-freerate A

5 10 CAPMwithzero-beta rate

10 5 CAPMwith risk-freerate

FIGURE PERFORMANCE CAPM AND OFTHE 3. ICAPM, 1963:7-2001:12 Notes: The four diagramscorrespondto (clockwise from the top left) the ICAPM with a free zero-betarate, the ICAPM with the zero-beta rate constrainedto the risk-free rate, the CAPM with a constrainedzero-beta rate, and the CAPM with an unconstrainedzero-betarate. The horizontalaxes correspondto the predictedaverage excess returnsand the vertical axes to the sample averagerealized excess returns.The predictedvalues are from regressionspresentedin Table 7. Trianglesdenote the 25 ME and BE/ME portfolios and asterisks the 20 risk-sortedportfolios.

in the first and 7.3 percentage points in the second subsample). C. Additional Robustness Checks We have performed a number of additional exercises to examine the robustness of our results. Full details are reportedin the Appendix, but we summarizethe results here. Bias.-Our asset pricing model Small-Sample relies on an estimatedvector autoregression that generates estimates of the two components of marketreturns.In small samples, our estimation methodology may yield biased estimates. Of

concern, persistentautoregressivecoparticular efficients may be biased downward (Maurice Kendall, 1954), and regression coefficients of returnson persistent forecasting variables may be biased upward (downward) if returns are negatively (positively) correlatedwith innovations to the forecasting variables (Stambaugh, 1999). One way to explore the effects of smallsample bias is to take the estimated VAR coefficients as the true data-generating process and We use these samgenerate repeated samples. ples to estimatenew VAR systems and calculate various statistics. The difference between the mean of these statistics and the statistic in the data-generating process is a measureof bias. Of

VOL.94 NO. 5

CAMPBELL AND VUOLTEENAHO: BAD BETA, GOOD BETA

1269

course, this measuredepends on the maintained data-generatingprocess, so it should be taken merely as indicative in small samples. Our findings about small-sample bias are as follows. First, there is only a negligible bias in the VAR coefficient of stock returns on the value spread, which is expected as innovations in the value spreadare almost uncorrelated with stock returns.Second, there is very little bias in the estimated volatilities of cash-flow and discount-ratenews. Third, in the function that maps the VAR shocks into news, there is an upwardbias in the negative coefficients of cashflow and discount-rate news on the value-spread shock. This upward bias makes the estimated coefficients closer to zero than the true coefficients, and thus understatesthe relevance of the value spreadfor the news terms. Fourth,all the biases together work to shrink the beta differences across growth and value stocks toward zero in the modem period. Fifth, there is some downward bias in the estimated premium for cash-flow beta in the moder period. Thus we conclude that small-sample bias works against us and bias corrections would most likely strengthenour results. Conditional Pricing.-One concern about our results might be that the test-asset betas and estimated preference parametersappear rather different in the first and second subsamples. Even if betas and the variance of the market returnhave changed over time, one would hope that the underlying preferences of investors have remained stable. One way to come at this issue is to estimate the preference parameters from a conditional model.
First, we estimate covariances
NCF,t + NCF,t-1)

returns on the fitted conditional covariances, imposing the restrictionsimplied by each asset pricing model. Allowing for continuous variationin the covariancesproducesthe following resultsthat are consistent with those reportedearlier. The risk premium on cash-flow covariance is much higher than that on discount-rate covariance. The implied risk aversion parameteris high but reasonable,with point estimates between 8 and 11. The two-beta model fits very well even with the ICAPM restrictions, while the CAPM fits poorly and is rejectedby the pricing errortests. We have also estimated the conditional pricing model for subperiods,allowing for different preference parameters. The two-beta model passes the asset-pricingtests with flying colors, while the CAPM performs poorly in the latter subsample. Furthermore,when using continuously time-varying betas and the covariance y formulation,the preferenceparameter appears quite stable across subsamples. Estimated y's range from 4 to 16 in the early sample and from 7 to 12 in the modern sample, depending on whether the zero-beta rate is assumed to equal the risk-free rate. Sensitivityto p.-An importantparameterin our model is p, the coefficient of loglinearization defined by Campbell and Shiller's (1988a) of approximation the log returnon an asset. We choose this parameterbased on a priori economic reasoning instead of estimating it from the data. Our robustness checks demonstratethat our main results are robust to reasonable variation in the parameterp. The value of p makes very little difference to any of the results in the early subsample. In the modern subsample the fit of the two-beta ICAPM is sensitive to p if the zero-betarateis restrictedto equal the Treasurybill rate,because then the zero-betarate and risk price of discount-ratebeta are both restrictedso changes in the estimate of discount-ratenews affect the fit of the model. The fit of the twobeta ICAPM is much less sensitive to p if the model allows a free zero-beta rate, for then it offsets changes in the estimate of discount-rate news with changes in the zero-beta rate. The model with free factor risk prices is very insensitive to p and always estimates a price of

and Cov(ri,t,-NDR,t-NDR, 1) for each test asset using a rolling 3-year (36month) window. We use these rolling covariance estimates as instrumentsthatpredictfuture covariances. Second, we regress the realized cross products (NcF,t + NcF,t -)ri,t and
(-NDR,t NDR,t

Covt(ri,t,

lagged (t - 2) rolling covarianceestimates and portfolio dummies in two pooled regressions. We define conditional covariances (CovDR and CoCF) as the fitted values of those regressions. Third, in period-by-period cross-sectional regressions, we regress the realized simple excess

)ri,t on the corresponding

1270

THEAMERICAN ECONOMICREVIEW

DECEMBER2004

cash-flow beta much higher than the price of discount-ratebeta. Data Frequency.-Although our main results are obtained from monthly data, we have repeated our tests with quarterlyand annualdata. Asset pricing tests are conducted only over the full sample for annual data, since subsample results are tenuous when we have lowerfrequency estimates of cash-flow and discountrate news. The results are consistent with the monthly results in that the estimatedpremiafor cash-flow betas are always higherthanthose for discount-ratebetas, althoughthe differences are smaller and less statisticallysignificantbecause they are estimated over the full sample period using low-frequency data ratherthan the modem subsample using high-frequency data. We have also performedthe subperiodexperiments for quarterly data. The subperiod point estimates obtained from quarterly data are very similar to those obtained from monthly data, although the results are noisy. Sensitivity to Additional State Variables.Our basic VAR includes the returnon a market stock index, the term spread, the smoothed price-earnings ratio, and the value spread. It omits two other variablesthat are often used to predict stock returns:the Treasury-billrate and the log dividend-priceratio. When we include these other variables in the VAR system, our main results are not materiallyaltered.We have also found that our results are robust to adding many otherknown returnpredictorsto the VAR system. It should be remembered,however, that our results depend critically on the inclusion of the small-stock value spreadin our aggregateVAR system. If we exclude this variablewe no longer find a large difference between the cash-flow betas of value stocks and growth stocks. D. Loose Ends and Future Directions A numberof unresolvedissues remain.First, we have used a model that assumes a constant variancefor the marketreturnand its two components. We can extend the model to allow for changing volatility of the marketreturn,in the mannerof Chen (2003), but in this case we must

measurenews aboutvolatility-adjusted discount rates rather than simply news about discount rates themselves. We believe that the properties of market discount-ratenews will be fairly insensitive to any volatility adjustment, since movements in market volatility appear to be relatively short-lived. Related to this, we can allow for dynamically changing betas rather than assuming, as we have done here, that betas are constantover long periods of time. Andrew Ang and Chen (2003) and Franzoni(2002) discuss alternativemethods for estimatingthe evolution of betas over time. We have assumed that the rationallong-term investor always holds a constant proportionof her assets in equities. But if expected returnson stocks vary over time while the risk-freeinterest rate and the volatility of the stock market are approximatelyconstant, the long-term investor has an obvious incentive to time the market strategically.In future work we plan to extend the model to examine whether a long-term investor who strategically allocates wealth into stocks and bonds would be better off overweighting small and value stocks than holding the stock portion of her portfolio at market weights. With this extension it will be important to handle changing volatility correctly, since a strategic market-timingportfolio will be heteroskedastic even if the stock marketportfolio is homoskedastic. We have nothingto say aboutthe profitability of momentum strategies (Ball and P. Brown, 1968; Narasimhan Jegadeesh and Titman, 1993). Although we have not examined this issue in detail, we are pessimistic about the two-beta model's ability to explain average returns on portfolios formed on past one-year stock returns or on recent earnings surprises. Stocks with positive past news and high shorttermexpected returnsare likely to have a higher fractionof theirbetas due to discount-rate betas, and thus are likely to have even lower return predictionsin the ICAPM than the already-toolow predictionsof the static CAPM. Our model is silent on what is the ultimate source of variationin the market'sdiscountrate. The mechanismthat causes the market'soverall valuation level to fluctuatewould have to meet at least two criteria to be compatible with our simple intertemporalasset-pricing model. The

VOL.94 NO. 5

CAMPBELL AND VUOLTEENAHO: BAD BETA, GOOD BETA

1271

shock to discount rates cannot be perfectly correlatedwith the shock to cash flows. Also, states of the world in which discount rates increase while expected cash flows remain constant should not be states in which marginal utility is unusually high for other reasons. If marginal utility is very high in those states, the discount-rate risk factor will have a high premium instead of the low premium we detect in the data. We have estimated the cash-flow and discount-ratebetas of value and growth stocks from the behavior of their returns, without showing how these betas are linked to the underlying cash flows of value and growth companies. Similar to our decomposition of the marketreturn,an individual firm's stock return can be split into cash-flow and discount-rate news. Through this decomposition, a stock's cash-flow and discount-rate betas can be further decomposed into two partseach, along the lines of Campbelland JianpingMei (1993) and Vuolteenaho (2002), and this decomposition might yield interestingadditionalinsights. Preliminary results in Campbellet al. (2003) suggest thatthe cash-flow propertiesof growth and value stocks are the main determinants theirbetas with the of cash-flow and discount-ratenews on the aggregate stock market.Ravi Bansal et al. (2003) also model the cash flows of value and growthstocks in relationto their risks in a consumption-based asset pricing model. Finally, our model has interesting implications for corporatefinance, specifically for the methodsused by corporations calculatea cost to of capital when evaluating investmentprojects. The two-beta model suggests that the most importantdeterminantof the cost of capital is not the market beta of a project but its cash-flow beta. This is consistent with the suggestion of William Brainardet al. (1991) that "fundamental beta" estimated from cash flows could improve the empiricalperformanceand usefulness of the CAPM. Cash-flow beta could be estimated using an econometric model, as we do here, but it is possible that simpler methods, such as estimating beta over long horizons or regressing returns on aggregate corporate profitability, would also provide useful estimates of cash-flow beta and thus of the cost of capital.

IV. Conclusions In his discussion of empirical evidence on market efficiency, Fama (1991) writes: "In the end, I think we can hope for a coherent story that (1) relates the cross-section properties of expected returns to the variation of expected returns through time, and (2) relates the behavior of expected returns to the real economy in a rather detailed way." In this paper, we have presented a model that meets the first of Fama's objectives and shows empirically that Merton's (1973) ICAPM helps to explain the cross-section of average stock returns. We propose a simple and intuitive two-beta model that captures a stock's risk in two risk loadings: cash-flow beta and discount-ratebeta. The returnon the marketportfolio can be split into two components,one reflectingnews about the market's future cash flows and the other reflecting news about the market's discount rates. A stock's cash-flow beta measures the stock's returncovariance with the former component and its discount-rate beta its return covariancewith the lattercomponent.Intertemporal asset pricing theory suggests that the "bad"cash-flow beta should have a higher price of risk than the "good"discount-ratebeta. Specifically, the ratio of the two risk prices equals the risk aversion coefficient that makes an investor contentto hold the aggregatemarket,and the "good"risk price should equal the variance of the returnon the market. Empirically, we find that value stocks and small stocks have considerably higher cashflow betas than growth stocks and large stocks, and this can explain their higher average returns.The post-1963 negative CAPM alphas of growth stocks are explained by the fact that their betas are predominantlyof the good variety. The model also explains why the sort on past CAPM betas induces a strong spread in average returnsduringthe pre-1963 sample but little spread during the post-1963 sample. The post-1963 CAPM beta sort induces a postranking spread only in the good discount-rate beta, which carries a low premium.Finally, the model achieves these successes with the discount-ratepremium constrained to the premodel. diction of the intertemporal

1272

THEAMERICAN ECONOMICREVIEW

DECEMBER2004

Our model has important implications for rational investors. While we do not show that such investors should hold the marketportfolio in preference to timing strategicallythe equity market,we do show that sufficientlyrisk-averse long-term investors who hold only equities should view the high average returnson value stocks and small stocks as appropriate compensation for risk rather than a justification for systematic tilts toward these types of stocks. Investors with lower risk aversion should overweight these stocks, while investors with higher risk aversion should underweight them. This analysis should be of interest even if one believes that investor irrationalityhas an important effect on stock prices, because even in this case one should want to know how rational investors perceive stock market risks. Our results have obvious relevance for such long-term institutionalinvestors as pension funds, which maintainstable allocations to equities and wish to assess the risks of portfolio tilts toward particular types of stocks. Ourtwo-beta model is, of course, not the first attempt to operationalize Merton's (1973) ICAPM. We hope, however, that our model is an improvementover previous specificationsin two respects. First, our specification"works"in the sense that it has respectable explanatory power in explainingthe cross-sectionof average asset returns with premia restricted to values predictedby the theory. Second, by restatingthe model in the simple two-beta form, with a close link to the static CAPM, we hope to facilitate the empirical implementation of the ICAPM in both academic research and practical applications. REFERENCES Adrian, Tobias and Franzoni, Francesco A. "Learningabout Beta: An Explanationof the Value Premium." HEC WorkingPaper,2004. Ang, Andrew and Chen, Joseph. "CAPM over the Long Run: 1926-2001." UnpublishedPaper, 2003. Ball, Ray. "Anomaliesin Relationshipsbetween Securities' Yields and Yield-Surrogates." Journal of Financial Economics, 1978, 6(2/ 3), pp. 103-26.

Ball, Ray and Brown,P. "An EmpiricalEvaluation of Accounting Income Numbers."Journal of Accounting Research, 1968, 6(2), pp. 159-78. Bansal,Ravi; Dittmar,RobertF. and Lundblad, ChristianT. "Interpreting Risk Premiaacross Size, Value, and IndustryPortfolios."Unpublished Paper, 2003. Banz, Rolf W. "The Relation between Return and MarketValue of Common Stocks."Journal of Financial Economics, 1981, 9(1), pp. 3-18. Performance Comof Basu,Sanjoy."Investment mon Stocks in Relation to Their PriceEarnings Ratios: A Test of the Efficient Market Hypothesis." Journal of Finance, 1977, 32(3), pp. 663-82. Basu, Sanjoy."The Relationshipbetween Earnings Yield, Market Value and Return for NYSE Common Stocks: FurtherEvidence." Journal of Financial Economics, 1983, 12(1), pp. 129-56. Black, Fischer. "Capital Market Equilibrium with RestrictedBorrowing."Journal of Business, 1972, 45(3), pp. 444-55. Brainard,William C.; Shapiro,MatthewD. and Shoven, John B. "FundamentalValue and Market Value," in W. C. Brainard,W. D. Nordhausand H. W. Watts,eds., Money,macroeconomics,and economicpolicy: Essays in honor of James Tobin. Cambridge,MA, and London:MIT Press, 1991, pp. 277-307.
Brav, Alon; Lehavy, Reuven and Michaely, Roni.

"Expected Return and Asset Pricing." Unpublished Paper, 2002.

Brennan, Michael J.; Wang, Ashley W. and Xia,

Yihong. "A Simple Model of Intertemporal CapitalAsset Pricing and Its Implicationsfor the Fama-FrenchThree-FactorModel." Unpublished Manuscript,2001.
Brennan, Michael J.; Wang, Ashley W. and Xia,

Yihong. "Estimation and Test of a Simple Model of IntertemporalCapital Asset Pricing." Journal of Finance, 2004, 59(4), pp. 1743-75. John Y. "StockReturnsand the Term Campbell, Structure." Journal of Financial Economics, 1987, 18(2), pp. 373-99.
Campbell, John Y. "A Variance Decomposition

for Stock Returns."Economic Journal, 1991, 101(405), pp. 157-79.

VOL.94 NO. 5

BAD BETA, GOOD BETA CAMPBELL AND VUOLTEENAHO:

1273

Asset Pricing Campbell,John Y. "Intertemporal without ConsumptionData."American Economic Review, 1993, 83(3), pp. 487-512. Risk and ReJohn Y. "Understanding Campbell, turn." Journal of Political Economy, 1996, 104(2), pp. 298-345.
Campbell, John Y. and Mei, Jianping. "Where

Questions."Journal of Business, 1999, 72(2), pp. 183-200.

D'Avolio, Eugene. "The Market for Borrowing

Stock." Journal of Financial Economics, 2002, 66(2-3), pp. 271-306.

Daniel, Kent and Titman, Sheridan. "Evidence

Do Betas Come From?Asset Price Dynamics and the Sources of Systematic Risk." Review of Financial Studies, 1993, 6(3), pp. 56792. teenaho,Tuomo."Growthor Glamour?"Unpublished Paper, 2003.

on the Characteristics of Cross Sectional Variation in Stock Returns."Journal of Finance, 1997, 52(1), pp. 1-33.
Davis, James L.; Fama, Eugene F. and French,

Campbell, John Y.; Polk, Christopher and Vuol-

Campbell, John Y. and Shiller, Robert J. "The

Dividend-Price Ratio and Expectations of FutureDividends and Discount Factors."Review of Financial Studies, 1988a, 1(3), pp. 195-228.
Campbell, John Y. and Shiller, Robert J. "Stock

Kenneth R. "Characteristics, Covariances, and Average Returns:1929 to 1997."Journal of Finance, 2000, 55(1), pp. 389-406. Dimson, Elroy. "Risk Measurement When Shares Are Subject to InfrequentTrading." Journal of Financial Economics, 1979, 7(2), pp. 197-226.
Eleswarapu, Venkat R. and Reinganum, Marc R.

Prices, Earnings, and Expected Dividends." Journal of Finance, 1988b, 43(3), pp. 66176. Campbell, John Y. and Shiller, Robert J. "Valu-

Epstein, Lawrence G. and Zin, Stanley E. "Sub-

"The Predictabilityof Aggregate Stock Market Returns: Evidence Based on Glamour Stocks." Journal of Business, 2004, 77(2), pp. 275-94.

ation Ratios and the Long-Run Stock Market Outlook."Journal of Portfolio Management, 1998, 24(2), pp. 11-26. Campbell,John Y. and Viceira, Luis M. "Consumption and Portfolio Decisions When Expected ReturnsAre Time Varying."Quarterly Journal of Economics, 1999, 114(2), pp. 433-95.
Chacko, George and Viceira, Luis M. "Dynamic

Consumptionand Portfolio Choice with Stochastic Volatility in Incomplete Markets." National Bureauof Economic Research,Inc., NBER Working Papers:7377, 1999. CAPM and the Chen, Joseph. "Intertemporal Cross-Section of Stock Returns." Unpublished Paper, 2003.
Cohen, Randolph B.; Polk, Christopher and Vuolteenaho, Tuomo. "The Price Is (Almost)

stitution, Risk Aversion, and the Temporal Behavior of Consumptionand Asset Returns: A Theoretical Framework." Econometrica, 1989, 57(4), pp. 937-69. Epstein,LawrenceG. and Zin, StanleyE. "Substitution, Risk Aversion, and the Temporal Behavior of Consumptionand Asset Returns: An EmpiricalAnalysis." Journal of Political Economy, 1991, 99(2), pp. 263-86. Fama,EugeneF. "EfficientCapitalMarkets:II." Journal of Finance, 1991, 46(5), pp. 1575617.
Fama, Eugene F. and French, Kenneth R. "Divi-

Fama, Eugene F. and French, Kenneth R. "Busi-

dend Yields and Expected Stock Returns." Journal of Financial Economics, 1988, 22(1), pp. 3-25. ness Conditions and Expected Returns on Stocks and Bonds." Journal of Financial Economics, 1989, 25(1), pp. 23-49. Cross-Section of Expected Stock Returns." Journal of Finance, 1992, 47(2), pp. 42765. mon Risk Factorsin the Returnson Stock and Bonds." Journal of Financial Economics, 1993, 33(1), pp. 3-56.

Right." National Bureau of Economic Research, Inc., NBER Working Papers: 10131, 2003a.
Cohen, Randolph B.; Polk, Christopher and Vuolteenaho, Tuomo. "The Value Spread."

Fama, Eugene F. and French, Kenneth R. "The

Journal of Finance, 2003b, 58(2), pp. 60941.

Cornell, Bradford. "Risk, Duration, and Capital

Fama, Eugene F. and French, Kenneth R. "Com-

Budgeting: New Evidence on Some Old

1274

THEAMERICAN ECONOMICREVIEW

DECEMBER2004

Franzoni, Francesco A. "Where Is Beta Going?

The Riskiness of Value and Small Stocks." UnpublishedPaper, 2004. Graham,Benjaminand Dodd, David L. Security analysis. New York: McGraw Hill, 1934. Guo, Hui. "Time-VaryingRisk Premia and the Cross Section of Stock Returns."FederalReserve Bank of St. Louis, Working Papers: 2002-013, 2002. Hodrick, Robert J.; Ng, David Tat-Chee and Sengmueller, Paul. "An International Dynamic Asset Pricing Model." International Tax and Public Finance, 1999, 6(4), pp. 597620. Hodrick,Robert J. and Zhang, Xiaoyan."Evaluating the Specification Errors of Asset Pricing Models." Journal of Financial Economics, 2001, 62(2), pp. 327-76. Jagannathan, Ravi and Wang, Zhenyu. "The ConditionalCAPM and the Cross-Section of Expected Returns." Journal of Finance, 1996, 51(1), pp. 3-53. Jegadeesh, Narasimhanand Titman, Sheridan. "Returns Buying Winnersand Selling Losto ers: Implications for Stock Market Efficiency." Journal of Finance, 1993, 48(1), pp. 65-91. RobertF. "PreKeim,DonaldB. and Stambaugh, dicting Returnsin the Stock and Bond Markets."Journal of Financial Economics, 1986, 17(2), pp. 357-90. Kendall,MauriceG. "Note on Bias in the Estimationof Autocorrelation." Biometrika,1954, 41(3-4), pp. 403-04. Kim, TongSuk and Omberg,Edward."Dynamic Nonmyopic Portfolio Behavior." Review of Financial Studies, 1996, 9(1), pp. 141-61. La Porta, Rafael. "Expectationsand the CrossSection of Returns." Journal of Finance, 1996, 51(5), pp. 1715-42. La Porta,Rafael;Lakonishok, Josef;Shleifer,Andrei and Vishny, RobertW. "Good News for Value Stocks: FurtherEvidence on Market Efficiency."Journal of Finance, 1997, 52(2), pp. 859-74. Lakonishok,Josef; Shleifer,Andrei and Vishny, Robert W. "Contrarian Investment,Extrapolation, and Risk." Journal of Finance, 1994, 49(5), pp. 1541-78. Lamont,Owen A. and Thaler,RichardH. "Can the MarketAdd and Subtract?Mispricing in

Tech Stock Carve-Outs." Journal of Political Economy, 2003, 111(2), pp. 227-68. Lettau, Martin and Ludvigson,Sydney. "Resurrecting the (C)CAPM: A Cross-Sectional Test When Risk PremiaAre Time-Varying." Journal of Political Economy, 2001, 109(6), pp. 1238-87. Lewellen,Jonathanand Nagel,Stefan."TheConditional CAPM Does Not Explain AssetPricing Anomalies." Unpublished Paper, 2003. Jonathanand Shanken, Lewellen, Jay. "Learning, Asset-Pricing Tests, and MarketEfficiency." Journal of Finance, 2002, 57(3), pp. 111345. Asset Pricing withLi, Yuming."Intertemporal out Consumption Data: Empirical Tests." Journal of Financial Research, 1997, 20(1), pp. 53-69. Liew, Jimmy and Vassalou,Maria. "Can Bookto-Market,Size and MomentumBe Risk Factors That Predict Economic Growth?" Journal of Financial Economics, 2000, 57(2), pp. 221-45. Lintner, John. "The Valuation of Risk Assets and the Selection of Risky Investments in Stock Portfolios and Capital Budgets." Review of Economics and Statistics, 1965, 47(1), pp. 221-45. Lo, Andrew W. and MacKinlay,A. Craig. "An Econometric Analysis of Nonsynchronous Trading." Journal of Econometrics, 1990, 45(1-2), pp. 181-211. Lynch, Anthony W. "Portfolio Choice and Equity Characteristics: Characterizing the Hedging Demands Induced by Return Predictability."Journal of Financial Economics, 2001, 62(1), pp. 67-130. McQueen,Grant;Pinegar,Michaeland Thorley, Steven."DelayedReactionto Good News and the Cross-Autocorrelationof Portfolio Returns."Journal of Finance, 1996, 51(3), pp. 889-919. Merton, Robert C. "An IntertemporalCapital Asset Pricing Model." Econometrica, 1973, 41(5), pp. 867-87. Mitchell,Mark;Pulvino,Toddand Stafford,Erik. Jour"LimitedArbitragein EquityMarkets." nal of Finance, 2002, 57(2), pp. 551-84. Ng, David Tat-Chee."The InternationalCAPM When Expected ReturnsAre Time-Varying."

VOL.94 NO. 5

CAMPBELL AND VUOLTEENAHO: BAD BETA, GOOD BETA

1275

Journal of InternationalMoney and Finance, 2004, 23(2), pp. 189-230. MiNg, Victor;Engle, RobertF. and Rothschild, chael. "A Multi-Dynamic-FactorModel for Stock Returns." Journal of Econometrics, 1992, 52(1-2), pp. 245-66. Allan. Perez-Quiros,Gabriel and Timmermann, "FirmSize and Cyclical Variations in Stock Returns."Journal of Finance, 2000, 55(3), pp. 1229-62. Peterson,James D. and Sanger,Gary C. "CrossAutocorrelations, Systematic Risk and the Period of Listing."UnpublishedPaper, 1995. Marc R. "Misspecificationof CapiReinganum, tal Asset Pricing: Empirical Anomalies Based on Earnings Yields and Market Values." Journal of Financial Economics, 1981, 9(1), pp. 19-46. Roll, Richard."A Critiqueof the Asset Pricing Theory's Tests: PartI: On Past and Potential Testabilityof the Theory."Journal of Financial Economics, 1977, 4(2), pp. 129-76. Rosenberg,Barr; Reid, Kenneth and Lanstein, Ronald. "PersuasiveEvidence of Market Inefficiency." Journal of Portfolio Management, 1985, 11(3), pp. 9-16. Ross,StephenA. "TheArbitrageTheoryof Capital Asset Pricing."Journal of Economic Theory, 1976, 13(3), pp. 341-60. Rozeff,MichaelS. "DividendYields Are Equity Risk Premiums."Journal of Portfolio Management, 1984, 11(1), pp. 68-75. Scholes, Myron and Williams,Joseph T. "Estimating Betas from Nonsynchronous Data."

Journal of Financial Economics, 1977, 5(3), pp. 309-27. Schroder, Mark and Skiadas, Costis. "Optimal Consumption and Portfolio Selection with Stochastic Differential Utility." Journal of Economic Theory, 1999, 89(1), pp. 68126. Shanken,Jay. "MultivariateProxies and Asset Pricing Relations: Living with the Roll Critique." Journal of Financial Economics, 1987, 18(1), pp. 91-110. Sharpe, William F. "Capital Asset Prices: A Theory of MarketEquilibriumunder Conditions of Risk." Journal of Finance, 1964, 19(3), pp. 425-42. Shiller,RobertJ. Irrationalexuberance.Princeton: PrincetonUniversity Press, 2000. RobertF. "On the Exclusion of AsStambaugh, Model: sets from Tests of the Two-Parameter A Sensitivity Analysis."Journal of Financial Economics, 1982, 10(3), pp. 237-68. RobertF. "PredictiveRegressions." Stambaugh, Journal of Financial Economics, 1999, 54(3), pp. 375-421. Vassalou,Maria. "News Related to FutureGDP Growth as a Risk Factor in Equity Returns." Journal of Financial Economics, 2003, 68(1), pp. 47-73. Tuomo. "WhatDrives Firm-Level Vuolteenaho, Stock Returns?"Journal of Finance, 2002, 57(1), pp. 233-64. Zhang, Lu and Petkova, Ralista. "Is Value Riskier Than Growth?"Unpublished Paper, 2002.